Classification of same limb motor imagery EEG using temporal attention based hierarchical transformer

Abstract

Applications of motor imagery (MI) based braincomputer interfaces (BCI) are frequently seen in medicine and robotics. Currently, most BCIs rely on distinct body parts such as the left hand, right hand, feet, and tongue. However, due to the limited number of independent control signals they provide, these are not ideal for complex system control. MI tasks within the same upper limb address this by enabling more intuitive system control, but have relatively few studies on them. It is challenging to classify these tasks because they activate closely spaced motor cortex regions. To address this, we propose an attention mechanism based hierarchical transformer architecture that selectively emphasizes temporal segments with important features by assigning higher attention weights. It consists of a low level transformer (LLT) layer that extracts features from short EEG segments, and a high level transformer (HLT) that uses selfattention to identify and combine key features for classification. The model achieved a competitive accuracy of 54.07% for four class same limb MI tasks, far surpassing the 25% chance level, and also demonstrated a reasonable robustness to session variability. Experimental results indicate the model’s effectiveness in classifying MI tasks of the same limb and potential for the advancement of BCIs.